InstrumentalEval: leaderboard
The model gets an ordinary goal such as making money, with a system prompt urging it to achieve the goal at all costs, in 76 tasks where avoiding shutdown, gaining unauthorised access, copying itself, hiding from monitoring or deceiving would help; a GPT-4o judge marks responses that pursue such a subgoal and the score is their share. Unlike shutdown-resistance tests in a sandbox, each task is a single written reply.
Metric: Responses pursuing an instrumental goal, all 76 tasks (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 21.05 | #741 |
| 2 | DeepSeek V3 | 21.05 | #716 |
| 3 | Gemini 2.0 Flash (01-21) (Thinking) | 32.89 | #711 (Gemini 2.0 Flash (01-21)) |
| 4 | O1 Mini | 43.42 | #764 |
| 5 | O1 (2024-12-17) | 44.74 | #311 |
| 6 | DeepSeek R1 | 46.05 | #473 |
| 7 | O3 Mini | 48.68 | #533 |
Interactive version: theaggregate.ai/benchmark?slug=instrumentaleval · How It Works · Data refreshed daily, snapshot 2026-11-13.